Mon.Not.R.Astron.Soc.000,1–17(2007) Printed2February2008 (MNLATEXstylefilev2.2) A robust statistical estimation of the basic parameters of single stellar populations. I. Method 1 2 X. Hernandez and David Valls–Gabaud 1InstitutodeAstronom´ıa,UniversidadNacionalAuto´nomadeMe´xicoA.P.70–264,Me´xico04510D.F.,Me´xico 8 2GEPI,CNRSUMR8111,ObservatoiredeParis,5,PlaceJulesJanssen,92195MeudonCedex,France 0 0 2 Accepted...Received...;inoriginalform... n a J ABSTRACT 9 Thecolour-magnitudediagramsofresolvedsinglestellarpopulations,suchasopenandglob- ularclusters,haveprovidedthebestnaturallaboratoriestoteststellarevolutiontheory.Whilst ] h a variety of techniqueshave been used to infer the basic propertiesof these simple popula- p tions,systematicuncertaintiesarisefromthepurelygeometricaldegeneracyproducedbythe - similarshapeofisochronesofdifferentagesandmetallicities.Herewe presentanobjective o r androbuststatistical techniquewhichlifts this degeneracyto a greatextentthroughthe use t ofakeyobservable:thenumberofstarsalongtheisochrone.ThroughextensiveMonteCarlo s a simulationsweshowthat,forinstance,wecaninferthefourmainparameters(age,metallicity, [ distanceandreddening)inanobjectiveway,alongwithrobustconfidenceintervalsandtheir fullcovariancematrix.Weshowthatsystematicuncertaintiesduetofieldcontamination,unre- 2 solvedbinaries,initialorpresent-daystellarmassfunctionareeithernegligibleorwellunder v 1 control.Thistechniqueprovides,forthefirsttime,aproperwaytoinferwithunprecedented 8 accuracy the fundamentalproperties of simple stellar populations, in an easy-to-implement 0 algorithm. 1 . Keywords: methods:statistical–stars:statistics–globularclusters:general–openclusters 1 andassociations:general–Galaxy:stellarcontent 0 8 0 : v 1 INTRODUCTION theotherhand,reliesonthedifferenceinapparentmagnitudebe- i X tweenhorizontalbranch(HB)starsandturn-offstars.Bothmeth- r Thetheoryofstellarevolutionhasbeentestedextensivelymainly odsrelyheavilyon“semi-empiricalcalibrations”providedbystel- a larmodels,andaresubjecttomanyuncertainties(e.g.,Howtode- throughwell-studiedbinarystars(e.g.,Popper1980,Lastennet& fineproperlytheTOpointwhenphotometricerrorsblurthisregion Valls-Gabaud 2002, Ribas 2006) or simple, coeval stellar popu- ofthecolour-magnitudediagram?WhatistheverticaloffsetHB– lationssuch asopen and globular clusters (e.g.Vandenberg et al. TOwhentheHBisnothorizontal?)whichneverthelesshavebeen 1996,Renzini&FusiPecci1988).Ithasbeenextraordinarilysuc- tackled withsome success (e.g.Rosenberg et al. 1999, Salaris& cessful at explaining most stellar properties across the different Weiss2002). evolutionarystages,byincorporatingincreasinglycomplexphysics in both stellar interiors and atmospheres (see Salaris & Cassisi Here we want to point out that many of the degeneracies 2006 for an updated and extensive review). In contrast, and per- areactuallymereartifacts,andthat aproper andrigorous mathe- hapssomewhatparadoxically,thedeterminationofthefundamen- maticaltechniquecanbeformulatedtomeasuretheseparameters talparametersofsinglestellarpopulations, suchasdistance,age, inarobust and objective way. For example, thewell-known age- metallicity,reddening,convectiveparameters,etc,hasseldombeen metallicitydegeneracy,whichstatesthatagivenisochronecanbe thesubject of arigorousmathematical determination. Moreoften fittedwithadifferentage providedthemetallicityischanged ap- thannot,isochronesare“fitted”byeye,andveryroughestimates propriately, isa purelygeometricaldegeneracy : theshapeof the oferrorsaregiven,ifgivenatall.Thereasonforthisstateofaffairs isochronesisthesame,yetstellarevolutionpredictsthatstarswith isperhapsthemanydegeneraciesbetweentheseparameters,which a different metal content will evolve at a differentspeed. Hence, makesitseeminglyimpossibletofindauniquesolutiontothesetof theage-metallicitydegeneracyisnotaphysicalone,butsimplya fundamentalparameters.Alternative,semi-empiricalmethodshave resultofusingthegeometricalshapeofisochronesastheonlydis- beendevelopedinsteadtoinferrelativequantities,ratherthanab- criminantfactor. Clearly,asimplestatisticalestimatorthatwould soluteones.Forinstance,theso-calledhorizontalmethodrelieson countthenumberofstarsalongtheisochroneswouldgiveverydif- the difference in colour between the turn-off (TO) point and the ferentresultsforanypairofage/metallicityvaluesthatwouldgive baseoftheredgiantbranch(RGB),andisindependent oftheas- the same geometrical shape to the isochrones. In a more formal sumptionsmadeontheconvectivetheory.Theverticalmethod,on way,letusconsideracurvilinearcoordinatesalonganisochrone 2 X. HernandezandD. Valls–Gabaud ofgivenparameters(say,distance,reddening,metallicity,age,al- detailsalong the RGB.Clearly, theoptimal strategy isto use the phaelementsenhancement,etc).Thispositiondependsonlyonthe fullinformationprovidedbytheobserveddensityofstarsalongthe agetofthatisochroneandonthemassmofthestaratthatprecise isochrone. locus,sothats=s(m,t)only.Anoffsetinagedtisreflectedonly throughchangesinmassmandpositionsalongtheisochroneof Tosummarizeexistingmethods,oneapproachhasgrownout agetsince of thefittingof ’geometrical’ indicators of theobserved CMD to isochrones,fromsingleparticularindicatorssuchasturnoffpoint ∂t ∂t dt(m,s) = dm + ds . (1) (TO), magnitude or colour difference between two points on the ∂m(cid:12)s ∂s(cid:12)m CMD, TO vs. zero-age horizontal branch or tip of the RGB and (cid:12) (cid:12) Forastarinthisis(cid:12)ochronetheoff(cid:12)setis,bydefinition,zero,hence RGB bump position and combinations thereof. Examples can be ∂m ∂m ∂s −1 found in Renzini & Fusi Pecci (1988), Buonanno et al. (1998), = − × . (2) Salaris&Cassisi(1998), Salaris&Weiss(1998), VandenBerg& ∂s (cid:12)(cid:12)t ∂t (cid:12)(cid:12)s (cid:16)∂t(cid:12)(cid:12)m(cid:17) Durrel (1990), Ferraro et al. (1999) and Rosenberg et al. (1999). The fi(cid:12)rst term is alw(cid:12)ays finite.(cid:12)The second term is the evolution- Theaccuracyoftheseapproacheshasbeenextendedbyincluding aryspeed,therateofchangeforagivenmassmofitscoordinate several such geometrical indicators simultaneously (e.g. Vanden- alongtheisochronewhentheagechangesbysomesmallamount. Berg 2000, Meissner & Weiss 2006), or a complete geometrical One can thinkof s asrepresenting some evolutionary phase, and comparisonbetweenobservedCMDandisochrone(e.g.Straniero so this term will be largewhen the phase is short-lived : a small & Chieffi 1991). The reduction of the CMD to a few numbers, variation in age yields a very large change in position along the given the varied and complex physics which enters in determin- isochrone.Alternatively,foragivenage,andsincethefirsttermis ingthemimpliesusingonlyasubsetoftheinformationavailablein alwaysfinite,awidevariation inpositionimpliesanarrow range theCMD,whiletheunavoidableobservationalerrorspresentnec- inmass.Thisisthecaseoftheredgiantbranchorthewhitedwarf essarilyleadtoambiguitiesintheidentificationofthekeypointsto coolingsequence1,forinstance.Ontheotherhand,slowlyevolving beused, whicharehardtoquantifyformally, andwhichresultin phasessuchasthemainsequencehavesmallevolutionaryspeeds uncertainconfidenceintervalsbeingassignedtotheinferencesde- and wide ranges in mass for a given interval along an isochrone. rived.Neglectingstellarevolutionaryeffectslimitstheinformation Clearly, themost important phases todiscriminatebetween alter- contentoftheisochronesagainstwhichtheCMDsarecompared. nativeagesandmetallicitieswillbethepostmain-sequenceones, wheretherangeofmassissmall(andhenceinsensitivetothede- Ontheotherhand,approachesconcentratingexplicitlyonthe tailsofthestellarmassfunction),andatthesametimewhereevo- stellar evolutionary effects have grown out of the already men- lutionaryspeeds arelarge.Ifweconsider themid-tolower main tioned analysis of luminosity functions presented by Paczyn´ski sequence,atfixedmetallicity,isochronesofallagestracethesame (1984), a projection of the CMD onto the vertical axis. Exten- locus, with only marginal changes in the density distribution of sionshavemostlycenteredontheeasytomeasureandagesensitive pointsamongstthem.Inthissense,oneoftheparameters,t,isto featuresoftheRGBluminosityfunction(e.g.Jimenez&Padoan, alargeextentabsentfromthemainsequence,whilephasesbeyond 1998,andZoccali&Piotto2000).Hereagain,observationalerrors arealwayssubstantiallyafunctionofbothageandmetallicity.The together withthe binning adopted in generating luminosity func- densityofstarsalonganisochroneistherefore tions, imply a limited use of the information available in the ob- dN dN ∂m ∂t servation,whilenotincludingtheinformationencodedbythege- = − × × . (3) ds (cid:16)dm(cid:17) (cid:16) ∂t (cid:12)s(cid:17) (cid:16)∂s(cid:12)m(cid:17) ometryofboththeCMDandisochroneslimitstheaccuracyofthe (cid:12) (cid:12) inferences. Thefirsttermisrelatedtoth(cid:12)einitialstel(cid:12)larmassfunction(IMF), and,asdiscussedabove,thesecondtermisfinitewhilethethirdisa There are have been a few attempts in the past at using the strongfunctionoftheevolutionaryspeed.Ifthemassaftertheturn- full information available inacolour-magnitude diagram, as pio- offisassumedtoberoughlyconstant,thisimpliesthattheratioin neered by Flannery & Johnson (1982) and further developed by the number of stars in two different evolutionary stages after the Wilson & Hurrey (2003) and Naylor & Jeffries (2006). Unfortu- turn-off will only depend on the ratio of their evolutionary time nately all these approaches, while mathematically correct, fail to scales2. includeproperlythekeydiscriminatingfactordiscussedabove,the Intermsofobservables,thiswell-knowrelationhasbeenex- stellar evolutionary speed, resulting in wide uncertainties in the ploited using luminosity functions as a proxy (as pioneered by inferred parameters. The only other attempt we are aware of, in Paczyn´ski 1984), that is, number counts of stars as a function of thecontext ofinferringparametersfromastellarpopulation, was apparent magnitude. This is not the isochrone coordinate s, but madebyJørgensen&Lindegren(2005),whoadaptedourformal- rather the projection of the isochrone on the vertical axis of the ism developed in a previous paper (Hernandez, Valls-Gabaud & colour-magnitudediagram.Thefactthatthesub-giantbranchoften Gilmore1999)toderivestellarages.Thepresentpaperisafurther appearsnearlyhorizontalinCMDsimpliesthatluminositybinsin androbustextensionofthisapproach.PaperIIinthisseries(Valls- this regime will not be discriminant. This is unfortunate because Gabaud&Hernandez2007)willapplythemethodtodifferentsets thisphaseismoreheavilypopulatedthantheredgiantbranch,and ofobservationsofstellarpopulations. thustheuseof luminosityfunctionsisnotoptimal. Likewise,the projection onthehorizontal axis, acolour, islesssensitivetothe In section 2 we present the rigorous probabilistic model, whichistestedwithsyntheticcasesinSection3.Insection4weex- ploreandquantifytheeffectsofsystematicuncertainties,insection 1 excludingthebottomofthecoolingsequence,wherewhitedwarfshaving 5wepresentanapplicationofthemethodtoarealcase,NGC3201. alargerangeofprogenitormassespileup 2 Inthecontextofstellarpopulation synthesis,thisisknownasthefuel Finallyinsection 6wepresent our conclusions on thisnovel ap- consumptiontheorem(Renzini&Buzzoni1983). proach. Robuststatisticalmeasures ofsinglestellarpopulations 3 2 PROBABILISTICMODEL where σ(L )2 and σ(C )2 are the errors which have to be ap- m m pliedtothetheoreticalcomparisonstarofmassmtomapinonto Intryingtorecoverthephysicalparametersofanobservedcluster, theobservedCMD.i.e.,eachtheoreticalstarisbeingthoughtofas webeginbyconstructingamodelwhereonly4numbersdetermine aGaussianprobabilitydensitydistributioncenteredon(L ,C ), theabove,theage,metallicity,distanceandreddeningoftheclus- m m withdispersionsσ(L )andσ(C ).Equation4ishenceequiva- ter,avector(T,Z,D,R).Henceforth,wewilldefineDasthetotal m m lenttotheconvolutionovertheentireCMDplaneoftwoGaussian verticaldisplacementofthestarsintheobservedCMDwithrespect probabilitydistributions,afirstonecomingfromthereportedob- totheir original positions.Thisverticaldisplacement isofcourse servedstar,andasecondonefromtheprobabilistictransferfunc- thedistancemodulusandthecontributionoftheextinctioninthe tionwhichwemustapplytoatheoreticalmodelstarcomingfrom observedband,thatis,D = µ+R × R,whereRisthecolour V anisochronetomapitontoaparticularobservedCMD. excessinthebandsconsideredandR theratioofselectivetototal V Ifwewantedtoconsidermodelstarsasmathematicalpointsin absorption.Notethatthesecontributionsare,ingeneral,correlated, thiscomparison,followingBayestheorem,wewouldidentifythe sinceatlargerdistancesthereddeningusuallyincreaseswithinthe probabilitythataparticularreportedrealstar,itselfa2-Dprobabil- Galaxy.NotealsothatR maydependonthelineofsightconsid- V itydistribution,wasinfactacertainmodelstar,withtheprobabil- ered,eventhoughauniversalextinctioncurveisusuallyassumed. itythatthisparticularmodelstarmightactuallycomefromthe2-D Forthesereasonsitissimpler(andmodel-independent)toconsider probability distribution of the reported star. To treat the problem a vertical offset, and a horizontal one, the colour excess, as two rigorously, we consider only the comparison between a reported independentquantities. starandamodelone,onlyafterthislastonehasbeenmappedonto Thus,weassumetheIMFtobeknown,thebinaryfractionto theplanewheretheobservedstarexists,throughaprobabilityfunc- be zero and the contamination of background/foreground starsto tion,asdescribedabove. benegligible.Thisconditionsareofcoursenottrueinarealcase, InEquation4anumericalnormalizationconstanthasbeenleft andwewillthereforetreatthemassystematics,theeffectsofwhich out,asinanycaselikelihood functionshavenoabsolute normal- weshallexploreindetailinsection4. izations and only relative values are meaningful. Whereas σ(L ) FromaBayesianperspective,givenanactualobservedCMD, i andσ(C )aresuppliedbyaparticularobservationofarealCMD, we want to identify the vector (T,Z,D,R) which has the high- i σ(L )andσ(C )havetobeestimated.Inapracticalapplication, est probability in having resulted in the actual observed cluster, m m onecouldobtaintheaverageofσ(L )andσ(C ),asfunctionsof as that from which the CMD most probably actually came from. i i magnitude,andusethosetwofunctionsasmodelsforσ(L )and Thismodel isessentially what was introduced and tested inHer- m σ(C ). nandez et al. (1999), and subsequently used in Hernandez et al. m Havingtheprobabilitythatasingleobservedstarisinfacta (2000a,2000b)toinferstarformationhistoriesofresolvedgalax- particularmodelstar,wecannowcomputetheprobabilitythatour ies,althoughmuchsimplifiedhereforthetreatmentofsinglestellar observedstarcamefromanypointalongaparticularisochronein populations. Wemustconstruct astatisticalmethodfordetermin- ourparameterspace(T,Z,D,R)as: ingwhatistheprobabilitythatanobservedCMDmighthavebeen theresultofagivenvector (T,Z,D,R).AnobservedCMDwill m1(T,Z) be defined by a set of i = 1 ··· N stars, each having a mea- ⋆ G (T,Z,D,R)= ρ(m;T,Z)G (T,Z,D,R,m)dm.(6) sured magnitude and colour (in whichever photometric bands or i Z i filters) (Li,Ci), with σ(Li) and σ(Ci) being the measured (de- m0(T,Z) ducedfromphotometrydatareductionprocedures)1-sigmaerrors In the above expression ρ(m;T,Z) is the density of stars in these quantities. Errors in magnitudes and colours will be as- alongtheisochroneofmetallicityZandageT,aroundtheteststar sumedasbeingGaussianinnature,anduncorrelatedtoeachother ofmassm.Asexplainedin§1,thisdensitydependsbothontheas- forthetimebeing.Hence,areportedstar(L ,C ;σ(L ),σ(C )), i i i i sumedIMFandontheevolutionaryspeedpredictedbythestellar willbethoughtofasatwo-dimensionalGaussianprobabilityden- evolutioncodeatthatpositionoftheCMD.Thedensityofpoints sitydistribution,centeredon(L ,C ),withdispersionsσ(L ),and i i i alongtheisochrone,withinacertainintervalaroundacertainmass, σ(C ). i inthehighlydiscriminatingregionbeyondthemainsequence,will We begin with the probability that a given observed star, of be determined essentially by the time spent by a star within that magnitudeandcolour(L ,C ),isinfactarealstarofsomeaget, i i interval,asgivenbythestellarevolutionarycodes.Ifalargepor- andsomemetallicity,Z,ofmassm,luminosityL andcolourC , m m tion of the main sequence is included in the study, the IMF will observedwithanassumedverticaloffsetofD,andareddeningof dominateρ(m;T,Z) overthatregion, asdiscussedin§1. Onthe R,bothinunitsofmagnitudes.Thiswillbegivenby otherhand,themassintervalbeyondtheturnoffpointissosmall, thatpracticallyanyconceivableIMFwillyieldidenticalresultsfor Gi(T,Z,D,R,m) = ρ(m;T,Z)inthisregion.TheRBGhasalowerdensityofpoints thanthemainsequencenotbecauseoftheIMF,butbecauseofthe 1 × faststellarevolutionaryspeeds. (cid:18)σL(m,i)σC(m,i)(cid:19) In practice, the extremely rapid variations in magnitude and × exp −DL2(m,i) ×exp −DC2(m,i) . (4) colour with mass over critical regions of the isochrones make it (cid:18) 2σ2(m,i) (cid:19) (cid:18) 2σ2(m,i) (cid:19) convenient to carry out the integration in Equation 6 over a pa- L C rameter measuring length over the isochrone, rather than mass, IntheaboveexpressionD (m,i)andD (m,i)arethedif- L C in which case ρ(m;T,Z) must include weighting factors due to ferencesinmagnitudeandcolourbetweentheithobservedstarand the relative duration of different stellar phases. Also, m (T,Z) agenericstarofage,metallicity,distancemodulus,reddeningand 0 and m (T,Z) are a minimum and a maximum mass considered mass(T,Z,D,R;m).Also, 1 alongeachisochrone,suchthattheluminositycompletenesslimit σ2(m,i)=σ(L )2+σ(L )2 , σ2(m,i)=σ(C )2+σ(C )2,(5) of the observations always corresponds to the luminosity of star L i m C i m 4 X. HernandezandD. Valls–Gabaud m (T,Z), while m (T,Z) corresponds to the star of age and whichsomeoftheoffspringsineachgenerationhavearandomly 0 1 metallicity (T,Z) at the tip of the RGB. The explicit exclusion selected“gene”switchedfrom“1”to“0”,orvice-versa.Thissim- of phases beyond thehelium flashreduces to acertain extent the plisticalgorithmhasbeenshowntoconvergeremarkablyefficiently resolutionofthemethod,butincreasessignificantlyitsrobustness, intotheglobalmaximumofcomplexmulti-dimensional surfaces, asitguarantees (seesection4) thatonly well-establishedphysics associatedtoverydifferentproblems.Amorecompletedescription areconsideredinthestellarevolutionarymodels. ofthealgorithmcanbefoundinCharbonneau(1995),withsuccess- We shall refer to G (T,Z,D,R) as the likelihood matrix, fulapplicationstodiverseastrophysicalproblemsbeingdescribed i since each element represents the probability that a given star, i, ine.g.Teriacaet.al(1999) forfittingradiativetransfermodelsto wasactuallyformedwithmetallicityZ,attimeT withanymass, solaratmospheredata,Sevensteretal.(1999)infittingparameters and that itwas observed witha distance modulus of D, and red- ofGalacticstructuremodelstoobservations,Metcalfe(2003),for deningofR.Thatis,inthisBayesianapproach,wetreatthestellar thefittingofwhitedwarfastroseismologyparametersorGeorgiev massmasanuisanceparameterwhichwemarginalizeover. &Hernandez(2005)intheproblemofinferringoptimalwindpa- The probability that an entire observed CMD with N stars rametersfromobservedspectrallineprofiles,amongmanyothers. ⋆ arosefromaparticularmodel,i.e.,fromaparticularchoiceofpa- rameters(T,Z,D,R),willnowbegivenby: N⋆ 3 TESTINGTHEMETHOD L(T,Z,D,R) = G (T,Z,D,R) . (7) i InthissectionwetestthemethoddescribedinSection§2,through Yi=1 theinversionofaseriesofsyntheticCMDs.Asinthiscasethepoint In this way, we have constructed an optimal merit function inparameterspacefromwhichaparticularCMDwassimulatedis which can be evaluated at any point within the relevant four di- known,wecanoptimallyassestheperformanceofthemethod. mensionalparameterspace(T,Z,D,R),givenanobservedCMD. Insimulatinganobserved CMDthefirststepistoobtainan Itisimportanttoremarkthatthismeritfunctionusesallinforma- adequateisochronelibrary.Differentgroupshavesucceededinpro- tionavailableintheCMD densely. Nobinning or averagingover ducingstate-of-the-artstellarevolutionarycodesusingmostly,but discreteregionsisnecessary,andnoaspectsofthedistributionof not all the time, the same input physics. Since systematic errors stars on the CMD are discarded. Also, all the rich stellar evolu- will be produced by the choice of the set of isochrones, we use tionaryphysicsisincludedinthecomparisonwiththemodels,as theresultsofthethreedifferentgroups(Bergbusch&VandenBerg theisochrones arenotbeing reducedtoadiscreteset ofnumbers 1997, VandenBerg et al. 2006, Demarque et al. 2004, Girardi et (e.g.thepositionoftheTO,RGBslope,etc.)orevencurvesonthe al.2002).Wehaveworkeddirectlywiththeoutputofstellarevo- CMD.Asweshallseeinsection§3,thefulluseofallfeaturesin lutionary codes to produce our own isochrone grid with a dense both data and models, allows aprecise estimation of the inferred grid including 180 time steps and 120 metallicity intervals. Each parameters. isochronecontainsapproximately300stellarmasses.Thetimeres- Following Bayes theorem, one would now want to identify olutionforagesofbetween4.5Gyrandtheupperlimitof18.0Gyr thatpointwhichmaximizesL,the4coordinateswhichmaximize isof0.2Gyr,withyoungeragesbeingsampledmoredensely.The theprobabilitythattheobservedCMDdidinfactarisefromapar- spacinginmetallicityislinearinthelogarithm,with0.03dexper ticularpointintheparameterspace,withthesetofphysicalcluster interval. Throughout the paper we shall identify metallicity with parameterswhichhavethehighestprobabilityofhavinggivenrise [Fe/H], when using theoretical isochrones, or comparing against totheobservedCMD. empirical inferences. The spacings of the isochrone grid will de- Severalmethodsexistforidentifyingtheglobalmaximumofa terminethemaximumresolutionofthemethod,independently of multidimensionalsurface.Moststandardmethodsrunintoconsid- whatapproachistakenatthefollowinglevelsoftheproblem.For erabledifficultiesincaseswherelocalmaximaexist,asourprob- thepurposes of presenting themethod, these gridsare more than lemturnsouttohave.Themostreliablewayofsolvingthisproblem appropriate. istoevaluateL throughoutadensegridinalloftheavailablepa- Itisimportanttonotethatthehighlynon-linearvariationsin rameterspace.Thishowever,ishighlytimeconsuming,andhence magnitudeandcolourwithmasswhichappearalongisochroneson itishighlydesirabletofindamoreefficientmethod. criticalregionsoftheCMDimpliesthatadirectinterpolationbe- Wehaveusedthegeneticalgorithmapproach,whereanumber tweentwoisochroneswillnotproducetheintermediaryone,butan of“agents”areplacedatrandompointswithinourparameterspace, averagecurvewhichbareslittleresemblancetotheoutputofastel- andthepositionofeachcodifiedintoabinarynumber,the“genes” larevolutionarycode.Inallcases,wehavelimitedtheisochronesto ofeachagent.Thenumberofbinarybitsusedforeachcoordinate belowtheheliumflash.Thisconditionrestrictstheinformationwe fixestheresolutionofthealgorithm,whichinthiscasecorresponds considerinourmethodtotheregionwheredisagreementbetween totheresolutionoftheisochronegridonthe(T,Z)plane,andto differentgroupsworkinginthefieldisataminimum. 0.01maginthe(D,R)plane.ThevalueofListhenevaluatedat Oncetheisochroneshavebeenestablished,anIMFischosen thepositionofeachagent,andtheresultisusedasafitnessfunc- todistributestarsamongstdifferentmasses.Thisfunctionprovesto tion.Asecondgenerationisthenconstructedby“mating”theorig- berelativelyunimportant,asthemassrangeofstarsoneisdealing inalagents,pairingthemandcombiningthe“genes”ofeachparent withissmallprovidedstarsareselectedfromonemagnitudeorso toproducetwo“offsprings”.Thefitnessfunctionisusedtoweight belowtheTOregionallthewaytothetipoftheRGB.Thisensures thematingprobability,suchthatthefitterindividualshaveagreater that the mass range probed by the observations is small. For ex- probabilityof“mating”.The“genes”oftheoffspringsareacombi- ample,inglobularclustersolderthan∼5Gyrthemassintervalis nationofthoseoftheirparents,andhenceasgenerationsproceed, around1M⊙andhenceinthisnarrowmassrangethedetailsofthe thepopulationgraduallymovesintoregionswherethefitnessfunc- IMFareirrelevant.Thedistributionofstarsalongtheisochroneis tionishigher,i.e.,intoregionsofhighL.Toavoidgettingtrapped amuchmoresensitivefunctionofthestellarevolutionarymodels, intolocalmaxima,aslight“mutationrate”isintroduced,through whichisinfactwhatdeterminesthatthemainsequenceishighly Robuststatisticalmeasures ofsinglestellarpopulations 5 13 13.5 14 14.5 -3 -2.5 0.05 0.1 0.1 -2.6 14.5 -2.8 0.08 14 -3 0.06 -3.2 13.5 -3.4 0.04 13 18.9 18.9 18.9 18.8 18.8 18.8 18.7 18.7 18.7 13 13.5 14 14.5 -3 -2.5 0.05 0.1 T(Gyr) log(Z) E(V-I) Figure1. (Left)SimulatedSingleStellarPopulationCMDhavinganageof13.9GyrandlogZ=−2.7.Thethick(black)curveshowstheinputisochrone, withtheoptimalfitisochronerecoveredbythemethodappearingasathin(red)curve.Dottedcurvesshowtheyoungestandoldestisochronesallowedbythe method,atthe1−σlevel,whichinfactspana1.0Gyrinterval.(Right) ThesixpanelsshowtheprojectionoftheerrorellipsoidresultingfromtheMonte CarlosimulationandinversionofaSSPwithage13.9GyrandmetallicitylogZ=−2.7,ontodifferentplanes.Thefilledcirclesshowinputparameters,and theemptyonestheaveragevaluesfortherecoveredparameters. populated,theredgiantbranchmuchmoresparselyso,andthatthe functionofthesimulatedCMDwithrespecttoanygivenpointon HRgapregionwillcontainveryfewstars.TheIMFweuseisthat our 4-Dparameter space. Asdiscussed insection §2,theseeking of Kroupa et al. (1993), although as mentioned above, any other oftheoptimalpointwillbecarriedoutthroughageneticalgorithm approximation to this function would yield the same results (see simulation, the workings of which, for this first example, we de- §4.1). scribebelow. HavinganisochronegridandanIMF,wecannowpickapoint The upper left panel of Figure 2 shows the evolution of the inour(T,Z,D,R)parameterspaceandplaceachosennumberof valueofthelikelihoodmeritfunctionof Equation7over thefirst starsintothemagnitude-colourplane.Forthefirstimplementation 100generationsofthegeneticsimulation.Thehorizontallinegives we have simulated an extreme example with close to 2000 stars thevalueofthissamefunction,evaluatedattheactualpointinour around point (13.9,−2.7,18.8,0.05), a single stellar population (T,Z,D,R)parameterspacewhichwasusedtoproducethesyn- of age 13.9 Gyr, metallicity −2.7 (that is, close to −1.0 in solar theticCMDbeinginverted.Thisfirstpanelshowsanincreasewith units,for asolar metallicityof logZ⊙ = −1.7), withadistance thegenerationnumberoftheoptimalvalueofthelikelihoodfunc- modulus of 18.8 magnitudes, and observed through a reddening tion, as the absolute best point found in all the simulation up to of 0.05 magnitudes. Next we model magnitude and colour errors a given generation is what is being plotted. This value increases as Gaussian with sigmas as appropriate for current HST or high veryrapidlyatfirst,astheinitiallyrandompointsbegintomigrate qualitylongintegrationground-based observations. Explicitlywe towardsregionsmorecloselyresemblingthesimulatedCMD,but used progress becomes increasingly slower. In fact, there is no defini- tivecriteriontoestablishwhenageneticalgorithmsimulationhas σ(L)=0.001×exp[(L+D−17.0)/2.25] finallyconverged,thishastobeinferredfromsimulationswithsyn- theticdataandrepeatedexperiments,asanapproximatenumberof σ(C)=0.001×exp[(L+D−17.0)/1.97] generations, suitableforaparticular problem, afterwhichnofur- whereListhemagnitude,intheV band,ofthetheoreticalstar,and therimprovementisachieved.Fortheproblembeingtreatedhere, C itsV −I colour.Atthispointwecannowproduceasynthetic wehavefoundthemethodstopsyieldinganyimprovementinthe CMD,andforthisfirsttestobtainthe2009pointsshowninFigure fitsin general well before 3500 generations. Wehave thus ended 1. A magnitude cut at V = 25 has been used, limiting analysis thegeneticsimulationsat3500generations.Thevalueofthelike- tostarsbrighter thanthiscutoff.Giventhehighly correlatedway lihoodfunctionevaluatedattheactualinputpointalwaysremains inwhich star populate the main sequence, together withgrowing abovetheonecorrespondingtotheoptimalonefoundbythesimu- and sometimes divergent errors towards higher magnitudes, it is lationforthisfirst100generationsshownhere. sufficienttorestrictanalysisto2magnitudesbelow the’turnoff’ The upper right panel of Figure 2 gives the evolution of the region.Inwhatfollowsweshallkeepthe25magnitudelowerlimit, optimaltimegridindex(TI)alongtheisochronegrid,foundupto although achanging magnitude cut asdescribed above yieldsthe acertainpoint.Inthiscase,theinputageof13.9Gyrcorresponds sameresults. toour timeindex 159 out of 180, over thisinitial100 generation Atthispoint,Equation7canbeusedtoobtainthelikelihood period the time index corresponding to the optimal model found 6 X. HernandezandD. Valls–Gabaud 0 20 40 60 80 1000 20 40 60 80 100 0 1000 2000 3000 0 1000 2000 3000 150 150 100 100 0 50 50 0 0 100 1 100 1 80 80 60 0 60 0 40 40 20 -1 20 -1 0 0 0 20 40 60 80 1000 20 40 60 80 100 0 1000 2000 3000 0 1000 2000 3000 Generation Generation Generation Generation Figure2. (Left) Thefourpanelsshow,inaclockwiseprogressionstartingattheupperleft,themaximumlikelihoodpointfoundbythemethod,theoptimal timeindex,theoptimaldistancemodulus,andtheoptimalmetallicityindex,allasfunctionsofgenerationnumberforthegeneticalgorithmsimulation,forthe first100generations.(Right) Sameastheleftpanels,butshowingthefull3500generationsofthegeneticalgorithmsimulation. fluctuates from an initial value of 120 to 178, from 6 to 18 Gyr, onetimegridintervalabove(0.2Gyr),11metallicitygridintervals andendsthisinitialperiod12gridpointsabovetheinputvalue,at below(0.4dex)andatadistancemodulusaround0.05magnitudes 16.9Gyr.Thebottomleftandrightpanelsshowthecorresponding largerthantheinputparameters. evolutionofthemetallicityindexandthedistancemodulus,again, Theaboveresultmakesitclearthatinferringstructuralparam- thehorizontallinegivestheinputparameters.Wecanseethatafter etersofstellarpopulations,eveninthesimplestcases,cannotbe 2generations theinitiallyrandom valueshaveapproached signif- attemptedwithoutafullconsiderationoftheprobabilisticnatureof icantly the input ones, and then show considerable variations of theproblem,andalso,cannotgiveerror-freeresults.However, it around 35 metallicity grid intervals (1.5 dex) and 0.8 mag in the ispossibletoevaluatetheintrinsicerrorsofthemethod,inconnec- distancemodulus. tiontoanyparticularimplementation,bytheMonteCarlomethod. Itisinterestingtore-plottheleftpanelof Figure2thistime Theexperiment isrepeated alargenumber of times,andthevar- overtheentire3500generationrange,thisisdoneintherightpanel ious answers analyzed to obtain mean inferred values with well ofFigure2wherewecanseethatsubstantialimprovementsinthe determineddispersionsandcorrelations. optimal likelihoodvaluefound arestilloccurring after1000 gen- This method naturally takes into consideration all internal erations.Astrikingfeatureseenintheupperleftpanelofthislast sources of error and dispersion inherent to the procedure being Figureisthatthegeneticalgorithmhasactuallyfoundanoptimal tested,andyieldsaccurateandreliableconfidenceintervalsonthe model which in fact yields a likelihood value above that of the inferences obtained. Figure1 (right) shows with filledcirclesthe input parameters. At around 1000 generations, theoptimal likeli- input parameters of the CMD of Figure 1 and with empty ones hoodfoundcrossesthehorizontallinegivingthelikelihoodvalue thecenterofthedistributionsoftherecoveredparametersafterthe ofEquation7oftheinputparameters,withrespecttothisparticu- wholeprocedureofconstructingasyntheticCMDandinvertingit larrealizationoftheCMDcorrespondingtothem.Thisapparently was repeated 20 times. The ellipses give projections of the error paradoxicalresultisinfactunavoidable,andhighlightsthestatisti- ellipsoidontothedifferentplanesbeingshown.Itcanbeseenthat calnatureoftheproblem.TheconstructionofaCMDfromapoint theonlysystematicwhichappearsinthiscasebeyondthe1σlevel inthesinglestellarpopulationparameterspacecannotbethought isan+0.03magnitudeoffsetintherecoveredvaluesoftheredden- ofasneitheradeterministicnorauniqueprocess.Evenbeforethe ingforthecluster.Allotherparametershavebeenrecoveredwith errorsassociatedwithprocuringtheobservationsenterthepicture, nosystematics,the1σerrorsforthis13.9Gyroldclusterbeing0.5 the process of sampling an IMF has already introduced a proba- Gyr in age, 0.38 dex in metallicity and 0.05 mag in the distance bilisticingredient intrinsictothecluster itself(e.g.Cervin˜oetal. modulus. 2002,Cervin˜o&Valls-Gabaud2003).Inthissense,obtainingasin- Theisochronecorresponding tothesolidcirclesisshownin glesimulatedCMDandusingittocomparewithanobservedone Figure1bythethicksolidcurve,whichpracticallyliesabovethe canonlybeatbestacrudefirstapproximationtotheinvertingofa thin(red)curveoftheisochronecorrespondingtotheemptycircles CMD.Inthisparticularcase,theconvolvingofthetheoreticalstars alongalltheCMD.Somedifferencesbetweenthetwoisochrones witherrorgaussianshasshiftedthepointsaroundinsuchawaythat are evident along the upper half of the red giant branch, where theisochrone(understoodasaseriesofstellarevolutionarymod- so few stars are expected, that none of the 2009 obtained in the els,notjustacurveonaCMD)whichhasthegreatestprobability CMD realization shown appear there. The dotted curves give the ofhavingyieldedtheCMDisnottheinputone,butonewhichis isochronescorrespondingtotheoldestandyoungestpointsonthe Robuststatisticalmeasures ofsinglestellarpopulations 7 11 12 -3.4 -3.2 -3 0.04 0.06 -3 0.07 12.5 -3.1 0.06 12 -3.2 0.05 11.5 -3.3 11 0.04 -3.4 10.5 -3.5 0.03 18.85 18.85 18.85 18.8 18.8 18.8 18.75 18.75 18.75 18.7 18.7 18.7 11 12 -3.4 -3.2 -3 0.04 0.06 T(Gyr) log(Z) E(V-I) Figure3. (Left)SimulatedSingleStellarPopulationCMDhavinganageof10.9GyrandlogZ=−3.05.Thethick(black)curveshowstheinputisochrone, withtheoptimalfitisochronerecoveredbythemethodappearingasathin(red)curve.Dottedcurvesshowtheyoungestandoldestisochronesallowedbythe method,atthe1σlevel,whichinfactspana1.48Gyrinterval.(Right) ThesixpanelsshowtheprojectionoftheerrorellipsoidresultingfromtheMonte Carlosimulationandinversionofasinglestellarpopulationwithage10.9GyrandmetallicitylogZ =−3.05,ontodifferentplanes.Thefilledcirclesshow inputparameters,andtheemptyonestheaveragevaluesfortherecoveredparameters. ellipseintheage-metallicityplane, whichareseentobracket the anextensiveregionofthe(T,Z)plane,merelyarelationbetween HR-gapandRGBregionsoftheCMDshown. thepossibleageandmetallicity,withnowayofrestrictingthean- The second example was constructed from point swertoanarrowerregion.Theabovehighlightsproblemsinherent (10.9,−3.05,18.8,0.05) in our parameter space, taken to to methods where information is being discarded, to which must represent atypical old SingleStellarPopulation (SSP).Intesting beadded ambiguitiesassociated tothedefinitionof the’turnoff’ themethodforsystematicsrelatedtoeffectsnotexplicitlyincluded point.Giventhecurrenthighaccuracyoftheobservationsandthe inthemodel,comparisonswillbemadeatthispointinparameter theoreticalisochrones,inspectionofsimulatedandobservedCMDs space,insection§4. makesitclearthatthereisinfactnosuchthingasaturnoff’point’, Figure3showstheCMDwhichresultsfromconvolving the butratheradiffuse,hazyregionwhichispronetosamplingfluctu- modeledstarswiththesamesimulatederrorsthanlasttime.Also ationsandthepresence,orotherwise,ofafewoddstars. shownaretheisochronefromwhichthesimulatedstarswerecon- Other correlations are evident in Figures1 and 3, for exam- structed,thick(black)curve,andtheoptimalonefoundbythege- pleaclearnegativecorrelationbetweentheinferredreddeningand neticalgorithm,thincontinuouscurve.Thedottedcurvesgivethe metallicity,alsoagenericfeatureofanystellarpopulationinversion oldest and youngest isochrones consistent with the Monte Carlo method,inherenttothestructureoftheisochrones.Itisinteresting simulationontheinvertingprocedureata1σlevel,thistimespan- tonote inthelower 3panels of Figures1 and 3that the inferred ning1.2Gyr.AnalogoustoFigure1,Figure3(right)givesthepro- distancetotheclustershowsnostrongcorrelationswithanyofthe jectionoftheerrorellipsoidfortheinferredparametersusingthe other3parameters,contrarytowhathappenswhentheisochrones MonteCarlomethod,projectedonto6differentplanes.Oneagain are reduced to a set of numbers such as location of the turn off seesthattherearenosystematicsbeyondthe1σlevel,fromwhich point,slopeoftheRGBetc. weconcludethatthemethodaccuratelyrecoverstheinputparam- Forathirdexample, intryingtocover points onour param- etersofthecluster,andyieldsmeaningfulconfidenceintervalsfor eterspacerepresentativeofoldSSPs,wesimulatedanold,metal theanswersupplied. poorsuchsystem.Startingfrompoint(13.9,−4.0,18.8,0.05)we LookingattheupperleftpanelofFigure3(right)weclearly construct a series of CMDs, one of which is shown in Figure 4, see the actual intrinsic age-metallicity degeneracy in the way the withcurvesanalogoustowhatappearsinFigures2and3.Itisim- 1σ ellipse in the (T,Z) plane slants from the upper left to the pressivetoseehownearlyindistinguishabletheshapesoftheinput lowerright.Optimalfitswithlargeagestendtocorrespondtolower andmeanrecoveredisochronesareontheCMD,togetherwiththe metallicities,andvice-versa.However,weseethathavingincluded isochronescorrespondingtotheoldestandyoungestisochroneal- alltheinformationavailablebothinthesimulatedCMDandinthe lowedbytheMonteCarlosimulationata1σ level.All4ofthese theoreticalisochrones minimizesthisdegeneracy andoneobtains curves appear practically on top of each other in this case, with closedconfidenceintervals.IftheCMDisreducedto,saytwonum- veryminordifferencesonlyapparentatthecurvaturebetweenthe bers,correspondingtothelocationofthe’turnoff’,whencompar- HRgapregionandtheRGB. ingtoasimilarlyrestrictiveview ofwhat thestellarevolutionary Theaboveisparticularlysurprisingwhenconsideringthatthe modelsencode,oneendsupwithaconfidenceregionwhichspans isochrones shown span almost a 1.5 Gyr-long time interval. Of 8 X. HernandezandD. Valls–Gabaud 12.5 1313.5 14 -4 -3.5 0.04 0.06 -3.2 0.07 14 -3.4 0.06 13.5 -3.6 0.05 -3.8 13 0.04 -4 12.5 -4.2 0.03 18.85 18.85 18.85 18.8 18.8 18.8 18.75 18.75 18.75 18.7 18.7 18.7 12.5 1313.5 14 -4 -3.5 0.04 0.06 T(Gyr) log(Z) E(V-I) Figure4. (Left) SimulatedSingleStellarPopulation(SSP)CMDhavinganageof13.9GyrandlogZ = −4.0.Thethick(black)curveshowstheinput isochrone,withtheoptimalfitisochronerecoveredbythemethodappearingasathin(red)curve.Dottedcurvesshowtheyoungestandoldestisochrones allowedbythemethod,atthe1σlevel,whichinfactspana1.44Gyrinterval.(Right) Thesixpanelsshowtheprojectionoftheerrorellipsoidresulting fromtheMonteCarlosimulationandinversionofaSSPwithage13.9GyrandmetallicitylogZ=−4.0,ontodifferentplanes.Thefilledcirclesshowinput parameters,andtheemptyonestheaveragevaluesfortherecoveredparameters. course, the other parameters besides the age have been adjusted stellar populations the mean recovered parameters appear within tomaximizethefitwhenshiftingtheage,inaccordance withthe onegridintervalinmetallicityoftheinputparameters,thetheoret- correlations shown inFigure 4, theoldest isochrone has thelow- icalmaximumresolutionoftheimplementationhasbeenrealized estmetallicitytogetherwithasomewhataboveaveragereddening attheseyoungages,evenafterhavingconsideredrealisticerrorsin correction. Thispoint again showsthat theisochrones havetobe magnitudesandcolours.Theageresolutionisstillabovethatofthe treatedascollectionsoffullstellarevolutionmodels,notreduced isochrones,butrobust1σerrorsofaround0.5Gyrareencouraging. toafewscalarshapeparameters,noteventosimplyshapesonthe Weaddthattheprecisionofthemethodinrecoveringdistance CMD. Havingused thefull shape of theisochrones continuously modulusandreddeningparametersissufficienttomakeitauseful throughoutthefullCMD,togetherwithinformationregardingthe toolinthedeterminationofdistancestosinglestellarpopulations, durationofdifferentstellarphasesinconstructingthemeritfunc- witherrorssometimesasgoodasafewhundredthsofamagnitude, tionofEquation7hasallowedamuchfinerdiscriminationwithin as seen in Figure 5. Again, the correlations persist, although the our(T,Z,D,R)parameterspace,muchbeyondwhatconsidering details of these correlations change from example to example in onlytheshapeoftheisochroneswouldhaveallowed.Changingthe waysthatreflectthedetailsoftheisochronesaroundeachregion, inputparametersoftheisochroneevenata2-3sigmalevel,would and which are dependent on both the magnitude of the assumed result in curves compatible with the shape of the simulated stel- observationalerrorsandtheactualnumbersofstarspresentinthe lardistribution,butwhichareruledoutduetohavingincludedall simulatedCMDs. availableinformationbothintheCMDandthestellarmodels. With simulated errors tending to zero the inferred parame- Again,nosystematicoffsetsbetweentheinputvaluesandthe ters would tend to the input ones, to within the resolution of the meanrecoveredonesappearsbeyondthe1σ level,ascanbeseen isochronegridused,always.Also,ifthestatisticalcomparisonbe- from Figure 4. In this same Figure we see the same correlations tween data and models is robust, the systematics and confidence evidentinthepreviousanalogousFigures,withtheconfidencein- intervalsshouldtendtozeroasthenumberofdatapointstendsto tervals,particularlyinageandmetallicityhavinggrownconsider- infinity,whichofcourseisneverthecase. ablyincomparisontoFigure3.Thisisobviouswhenoneconsid- ershow consecutive isochrones equally spaced intimewillshow progressivelysmallerdifferencesastheageofastellarpopulation increases.Havinglefttheassumedobservationalerrorsandstellar 4 SYSTEMATICS numbersconstant,theuncertaintiesintheinferredparametershave growningoingtoanold,metalpoorpopulation. Havingtestedsuccessfullythemethodacrossrepresentativepoints Forthefourthandfifthexampleswecompleteoursamplingof inthe(T,Z,D,R)space,wenowturntotheproblemofestimating SSPsparameterspacewithtwoclustersat(6.9,−2.0,18.8,0.05) howrobusttheinferencesofourmethodwillbeinmorerealistic and(6.9,−3.37,18.8,0.05).Figures5and6arecompletelyanalo- cases, where physical ingredients not explicitlyconsidered in the goustothecorrespondingpreviousfiguresandessentiallyillustrate statistical modeling behind Equation 7, and present in a real ap- the same points mentioned before. For these younger simulated plication, will complicate matters, and might result in systematic Robuststatisticalmeasures ofsinglestellarpopulations 9 6.5 7 -2.1 -2 -1.9 0.02 0.04 0.06 -1.9 0.07 7.4 0.06 7.2 -1.95 0.05 7 -2 0.04 6.8 -2.05 0.03 6.6 -2.1 0.02 6.4 18.85 18.85 18.85 18.8 18.8 18.8 18.75 18.75 18.75 18.7 18.7 18.7 6.5 7 -2.1 -2 -1.9 0.02 0.04 0.06 T(Gyr) log(Z) E(V-I) Figure5. (Left)SimulatedSingleStellarPopulationCMDhavinganageof6.9GyrandlogZ=−2.05.Thethick(black)curveshowstheinputisochrone, withtheoptimalfitisochronerecoveredbythemethodappearingasathin(red)curve.Dottedcurvesshowtheyoungestandoldestisochronesallowedbythe method,atthe1σlevel,whichinfactspana1.0Gyrinterval.(Right)ThesixpanelsshowtheprojectionoftheerrorellipsoidresultingfromtheMonteCarlo simulationandinversionofaSSPwithage6.9GyrandmetallicitylogZ =−2.05,ontodifferentplanes.Thefilledcirclesshowinputparameters,andthe emptyonestheaveragevaluesfortherecoveredparameters. offsetsbetweentheactualphysicalparametersofastellarpopula- oftheMonteCarloprocedurecanbeseeninFigure7,whichfrom tionandourinferences. the previous discussions would not have been expected to show Inthissectionweexploresuchcasesbyconstructingsynthetic any substantial difference to Figure 3, as is indeed the case. By CMDsunderassumptionsdifferenttotheonesthemethodusesin comparingwithFigure3wenotenosubstantialdifferences,allin- theinversionprocedure,andcomparehowserioustheseeffectsare ferredparametersarewellwithinthe1σerrorestimates,withvery inalteringourinferences.The5effectsconsideredwillbe:(1)Er- lowuncertaintiesofaround0.5Gyrinage,0.2dexinmetallicity, rorsintheassumedIMF,(2) thepresenceofunresolved binaries, 0.01and0.05maginreddeninganddistancemodulus,respectively. (3)theeffectsofreducingthenumbersofstarspresentintheob- Theisochronescorrespondingtotheinput,inferred,andoldestand servedCMD,(4)systematicsassociatedwiththedifferentsetsof youngestvaluesrecovered(ata1σlevel),practicallyoverlapover stellartracksusedinconstructingtheisochronegrids,and(5)the theentireregionoverwhichsignificantnumbersofstarsarefound. presenceofcontaminatingstarsintheobservedCMDs. Weseethatthetheoreticalexpectationsoftheprevioussectionsare borne out, and our results are in fact highly robust to even large scalefeaturesoftheIMF.Thisisreassuringfromthepointofview 4.1 EffectsoftheIMF ofinferringages,metallicities,distancesandreddeningcorrections Figure7givesaCMDresultingfromthesamepointinparameter ofresolvedstellarpopulations,butdishearteningfromthatofinfer- spaceasFigure3,(10.9,−3.05,18.8,0.05),butcreatedthistime ringtheIMFofobservedstellarclustersthroughmethodssimilar usingasubstantiallydifferentIMF,heavilyweightedtowardsmas- towhathasbeendevelopedhere. sive stars, increasing the slope with respect to the Kroupa et al. (1993)IMFsuchthatthetypicalstellarmassincreasesbyafactor 4.2 Effectsofunresolvedbinarystars of3.Obtainingthesamenumberofpointsthusrequiresturninga much largertotal massintostars, butunder theassumption of an Thesecondtestwasperformedagainusingthesameisochroneli- equalnumberofstarsasinFigure3,theresultsarehighlysimilar. brary, and the input point (10.9,−3.05,18.8,0.05) of Figure 3, Theprecisemanner inwhichthedensityofstarsvariesalongthe butthistimetooneintwoofthesimulatedstarsweaddedanun- mainsequenceregionissomewhatdifferentfromthecaseofFig- resolvedcompanion,chosenfromthesameIMF,i.e.,thesynthetic ure3,butgiventhehighlycorrelatedmannerinwhichmagnitudes CMDwasproducedassuminga50%unresolvedbinarycontribu- andcoloursscalealongthisregion,theinformationcontentofthe tion.Theinversionprocedurewashoweverexactlythesameasin CMDisonlymarginallyaltered.Thehighlydeterminant sections thepreviouscases,withEquation7remainingunchanged,abinary beyond the main sequence span such a small mass range that all fractionof0%wasassumedinrecoveringtheinferredparameters. ofthesefeaturesarevirtuallyleftunchangedbyeventhestrongest OneoftheresultingCMDswhichwasusedintheMonteCarlo modificationsintheIMF. procedure for the inversion is shown in Figure 8, together with Theinversionprocedurewasrepeated20times,butassuming isochronesasinthepreviousFigures.Thepresenceof50%bina- thesameIMF asinFigure3, whichinthiscasenolonger corre- riesisparticularlyconspicuousalongthemainsequence,wherethe spondedtotheIMFusedtocreatethesyntheticCMDs.Theresults distributionofpointsisclearlysmearedtowardsreddercolours,ev- 10 X. HernandezandD. Valls–Gabaud 6.5 7 7.5 -3.6 -3.4 -3.2 0.02 0.04 0.06 -3.2 0.07 0.06 7.5 -3.3 0.05 -3.4 0.04 7 -3.5 0.03 6.5 -3.6 0.02 18.85 18.85 18.85 18.8 18.8 18.8 18.75 18.75 18.75 18.7 18.7 18.7 6.5 7 7.5 -3.6 -3.4 -3.2 0.02 0.04 0.06 T(Gyr) log(Z) E(V-I) Figure6. (Left)SimulatedSingleStellarPopulationCMDhavinganageof6.9GyrandlogZ=−3.37.Thethick(black)curveshowstheinputisochrone, withtheoptimalfitisochronerecoveredbythemethodappearingasathin(red)curve.Dottedcurvesshowtheyoungestandoldestisochronesallowedbythe method,ata1σlevel,whichinfactspana1.0Gyrinterval.(Right) ThesixpanelsshowtheprojectionoftheerrorellipsoidresultingfromtheMonteCarlo simulationandinversionofaSSPwithage6.9GyrandmetallicitylogZ =−3.37,ontodifferentplanes.Thefilledcirclesshowinputparameters,andthe emptyonestheaveragevaluesfortherecoveredparameters. identwhencomparing,forexample,withFigure7.Afewbinaries ever,havingweightedalltheCMDequally,themuchlargermain canalsobeseenaroundthecriticalturnoffregion,butthesehave sequence population, and the phases beyond the turn off, where hadlittleeffect.Giventhelargedifferenceinluminositybetween errorsare much smaller, anchor to alarge extent the fit against a aRGBstarandanymainsequenceone(themostlikelysecondary smallnumber of bluestragglersorbinariesveryclosetotheturn tobeaddedtoabinary),binarypollutionalongtheRGBproduces off.Morethan2localerrorsigmaawayfromtheturnoff,youcan absolutelynoeffects. fillinalmostanynumber ofbluestragglers,byconstructiontheir Bylookingatthebestfitinferredisochronesatthebaseofthe effectwillbenegligible,seeEq.4. main sequence, it can be seen that these have all shifted towards Inthe context of invertingresolved stellar populations, abi- theright,inresponsetotheweightingofthedistributionofpoints narydoesnotnecessarilyimplyaphysicalgravitationalassociation towards this direction. By looking at Figure 8 one can see that betweentwostars,butmerelythatthelightoftwoindividualstars thisshifttowardstherighthasbeenachievedoptimallybythege- has appeared in thedata as coming from asingle point. The two neticalgorithmnotbyalteringthereddeningcorrection,butrather starshenceshowasasingleonehavingthetotalluminosityofthe through an offset in the distance modulus. Whilst inferred age, sumforthetwostars,andaneffectivetemperaturebeingtheenergy metallicityandreddeningparametershavebeenobtainedwithinthe weightedsumofthetemperaturesofthetwostars,asperaStefan- 1σ errorellipses,theinferreddistancemodulusisoffatarounda Boltzmannlaw.Therefore,especiallyifthinkingofdensesystems 2σlevel,whichhowevercorrespondsonlyto0.15magnitudes.The suchasglobularclusterswherecrowdingeffectscanbedominant, clearest correlation between theinferredparameters isstillinthe theadoptionofthesameIMFforthesecondarycomponentofthe age-metallicityplane,andremainsinthesamesenseasexpected. binaryasfortheprimaryoneisjustified. Otherthantheslightoffsetininferreddistancemodulus,the onlyothereffectofhavingintroducedarelativelyhighbinarypopu- 4.3 Effectsofdifferentsetsofisochrones lationhasbeenthebroadeningoftheerrors.Theconfidenceregion isthistime2.5Gyracross,comparedto1.4GyrforFigure3.Again, Wenowturntoapossiblyimportantsystematic,relatedtotheva- theisochronesinFigure8areremarkablyclosetoeachother,but lidityof the underlying hypothesis that a single real star from an althoughregionsbeyondtheturnoffhavenotbeenmuchaffected observed CMD doesinfact behaveand evolveasthestellarevo- bytheinclusionofthebinaries,thefactthatthemainsequencehas lutionarycodeoneisusingtoconstructanisochronegridassumes shiftedoutofcorrespondencewiththeseotherregionshasresulted itto.Figure9givestheCMDproducedagainfromthesamepoint inpoorerfits,andalthoughnosystematicoffsetsinage,metallic- (10.9,−3.05,18.8,0.05) in parameter space, but using this time ityor reddening correction have appeared, theerror ellipseshave thePadovastellarevolutionarymodels(Girardietal.2002),whilst certainlyincreasedinsize. inallpreviousexamples,theYaleisochroneswhereused(Demar- Theparticularcaseofbluestragglershasnotbeentreated,but queetal.2004).Thedistributionofstarsonthemagnitude-colour a good approximation can be obtained from the binary contami- planeismuchthesameasinFigure3,andnosignificantdifferences nation example included. Binaries and blue stragglers can popu- canbedetected. latetheregionclosetotheturnoff,andresultinpoorerfits.How- AnumberofsuchCMDswheretheninvertedusingtheYale